Do we know for certain that the "watermark" is some hidden, yet statistically detectable, word or syntax choices? Because if AI is as good as Anthropic says, watermarking might not be explicable to baseline humans with small short term memory, 5+/-2 items of working memory, and loads of cognitive biases. Anthropic's AI might have some method that just looks like soothsaying to us.
Other than that, I do think Gruber is correct in this case.
Do we know for certain that the "watermark" is some hidden, yet statistically detectable, word or syntax choices? Because if AI is as good as Anthropic says, watermarking might not be explicable to baseline humans with small short term memory, 5+/-2 items of working memory, and loads of cognitive biases. Anthropic's AI might have some method that just looks like soothsaying to us.
Other than that, I do think Gruber is correct in this case.